Research map / Reinforcement learning / RL from feedback
WorldCompass: Reinforcement Learning for Long-Horizon World Models
Zehan Wang, Tengfei Wang, Haiyu Zhang, Xuhui Zuo, Junta Wu, Haoyuan Wang, Wenqiang Sun, Zhenwei Wang, Chenjie Cao, Hengshuang Zhao, Chunchao Guo, Zhou Zhao
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply WorldCompass RL post-training | HY-Video-1.5 · WorldPlay long-term, Combined Action | action accuracy: 19.70 -> 54.82 | Improved |
| apply WorldCompass RL post-training | Wan2.2 · WorldPlay long-term, Combined Action | action accuracy: 19.58 -> 42.92 | Improved |
| apply WorldCompass RL post-training | HY-Video-1.5 · WorldPlay short-term, Combined Action | action accuracy: 21.74 -> 58.20 | Improved |
| apply WorldCompass RL post-training | HY-Video-1.5 · WorldPlay long-term, Basic Action | action accuracy: 64.28 -> 76.56 | Improved |
| apply WorldCompass RL post-training | HY-Video-1.5 · WorldPlay long-term, Combined Action | HPSv3: -0.33 -> 0.73 | Improved |
| use clip-level instead of sample-level rollout (ablation) | WorldCompass (HY-Video-1.5) · WorldPlay long-term, Combined Action | action accuracy: 12.45 -> 54.82 | Improved |
| add visual quality reward (ablation) | WorldCompass (HY-Video-1.5) · WorldPlay long-term, Combined Action | HPSv3: -2.67 -> 0.73 | Improved |
| add visual quality reward (ablation) | WorldCompass (HY-Video-1.5) · WorldPlay long-term, Combined Action | action accuracy: 36.39 -> 54.82 | Improved |
| add interaction following reward (ablation) | WorldCompass (HY-Video-1.5) · WorldPlay long-term, Combined Action | action accuracy: 11.51 -> 54.82 | Improved |
| add interaction following reward (ablation) | WorldCompass (HY-Video-1.5) · WorldPlay long-term, Combined Action | HPSv3: 1.01 -> 0.73 | Hurt |
| use DiffusionNFT instead of DanceGRPO (ablation) | WorldCompass (HY-Video-1.5) · WorldPlay long-term, Combined Action | action accuracy: 20.02 -> 54.82 | Improved |
| add best-of-N sampling (ablation) | WorldCompass (HY-Video-1.5) · WorldPlay long-term, Combined Action | action accuracy: 55.28 -> 54.82 | No clear effect |
Closest papers by meaning
- RLVR-World: Training World Models with Reinforcement Learning (NeurIPS 2025 · similarity 0.90)
- Scaling Offline Model-Based RL via Jointly-Optimized World-Action Model Pretraining (ICLR 2025 · similarity 0.88)
- Zero-shot World Models via Search in Memory (NeurIPS 2025 · similarity 0.88)
- WorldReasonBench: Human-Aligned Stress Testing of Video Generators as Future World-State Predictors (NeurIPS 2026 · similarity 0.88)
- Reinforcing VLAs in Task-Agnostic World Models (NeurIPS 2026 · similarity 0.87)
- Open-World Reinforcement Learning over Long Short-Term Imagination (ICLR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
